toto-2 2.0.0

Toto 2.0 — foundation model for multivariate time series forecasting

Toto 2.0

Technical Report | Blog | Model Collection | BOOM Dataset | GitHub

Toto 2.0 is a family of foundation models for multivariate time series forecasting, built by Datadog. It features a u-μP-scaled transformer with alternating time/variate attention and quantile-based probabilistic forecasting, ranging from 4M to 2.5B parameters.

Note: Fine-tuning and exogenous variable support are planned for a future 2.0 release. If you need these features today, see toto-ts (Toto 1.0).

Features

• Zero-Shot Forecasting — No task-specific fine-tuning required.
• State-of-the-Art Performance — Top results on GIFT-Eval and BOOM.
• Multivariate Support — Efficiently handles multiple variables via alternating time/variate attention.
• Probabilistic Predictions — Returns 9 quantile levels (0.1–0.9) for uncertainty estimation.
• High-Dimensional Support — Scales to time series with a large number of variates.
• Decoder-Only Architecture — Supports variable prediction horizons and context lengths.

Model Weights

Checkpoint	Parameters
Toto-2.0-4m	4M
Toto-2.0-22m	22M
Toto-2.0-313m	313M
Toto-2.0-1B	1B
Toto-2.0-2.5B	2.5B

Installation

Requires Python 3.12+ and PyTorch 2.5+. A CUDA-capable GPU (Ampere or newer) is recommended.

pip install toto-2

Or install the toto-models umbrella package, which includes toto-2 and its dependencies:

pip install toto-models

Quick Start

import torch
from toto2 import Toto2Model

model = Toto2Model.from_pretrained("Datadog/Toto-2.0-22m")
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
model = model.to(device).eval()

# Input shape: (batch, n_variates, time_steps)
target = torch.randn(1, 1, 512, device=device)
target_mask = torch.ones_like(target, dtype=torch.bool)
series_ids = torch.zeros(1, 1, dtype=torch.long, device=device)

# Returns quantiles of shape (9, batch, n_variates, horizon)
# Quantile levels: [0.1, 0.2, 0.3, 0.4, 0.5, 0.6, 0.7, 0.8, 0.9]
quantiles = model.forecast(
    {"target": target, "target_mask": target_mask, "series_ids": series_ids},
    horizon=96,
    decode_block_size=768,   # None for single forward pass (faster, better short-term)
    has_missing_values=False, # Set True if target_mask contains False entries
)

Inference tips:

• decode_block_size=None — single forward pass, faster and better for short horizons. Used for all leaderboard results.
• decode_block_size=768 — block decoding, better long-term stability for horizons ≳1000. Default in notebooks.
• has_missing_values=False — enables Flash Attention kernels when your context has no gaps.

Tutorials

• Quick Start — Load a model, forecast, plot results, handle missing values and multivariate inputs.
• GluonTS Integration — Use Toto2GluonTSModel with GluonTS evaluation pipelines and built-in datasets.

Evaluation

• GIFT-Eval Notebook — Evaluate on the GIFT-Eval benchmark.
• BOOM Evaluation — Evaluate on the BOOM observability benchmark.

Citation

If you use Toto 2.0 in your research, please cite:

@misc{khwaja2026toto20timeseries,
      title={Toto 2.0: Time Series Forecasting Enters the Scaling Era},
      author={Emaad Khwaja and Chris Lettieri and Gerald Woo and Eden Belouadah and Marc Cenac and Guillaume Jarry and Enguerrand Paquin and Xunyi Zhao and Viktoriya Zhukov and Othmane Abou-Amal and Chenghao Liu and Ameet Talwalkar and David Asker},
      year={2026},
      eprint={2605.20119},
      archivePrefix={arXiv},
      primaryClass={cs.LG},
      url={https://arxiv.org/abs/2605.20119},
}

License

Apache-2.0. See LICENSE for details.